In the leadup to Dreamforce 2026, dozens of announcements outlined what enterprises look like when software meets AI, but the event itself filled in the picture. Organizations are being transformed. Ways of working within them are transforming just as quickly. To see how headless architecture, Slack Code, purpose-trained models, and more are going to shape the future, read on for the most compelling takeaways from Dreamforce 2026.

1. Software used to make you come to it. Now it comes to you.

The way people interact with enterprise applications hasn't really changed for decades, said Khushwant Singh, SVP of Product Management for Platform, at the AIforce Keynote. "You open an application, you navigate to the right page, you complete the task … and you do that all over again."

AI inverts that: Agents don't need a screen. They need the capabilities beneath the screen - the APIs, metadata, business semantics, permissions, and governance that Salesforce has spent 27 years building. And now with AIforce, all those capabilities are available to agents through any AI interface - no need to log into Salesforce at all. "Agents need a different entrance into the enterprise," Singh said. "That's why we opened up Salesforce and all of its trusted capabilities to any AI, any interface, any agent, without requiring a Salesforce UI."

"Other software companies, they think their product is the UI," Patrick Stokes, Salesforce President of Applications and Marketing, said during the Main Keynote. "At Salesforce, our product is the trust that all of you, our customers, put into us to hold your data, to hold your workflows, your business processes, your permissions, your security rules. That is the Salesforce product."

Giving agents direct access to the information and metadata inside Salesforce unlocks tremendous value and powerful new ways of working. For one thing, you no longer have to know how to use Salesforce to use Salesforce. Building an app or updating a record used to require specialized training. Now, it's available to anyone via natural-language prompt, making the power of the platform accessible to a new generation of builders.

And those builders can now tap into the entire platform, pull out the relevant information and functionality, and create custom interfaces that give them everything they need and nothing they don't. That's because AIforce turns all the metadata that admins and developers have encoded inside Salesforce over the last 27 years - the definitions of objects, fields, relationships, permissions, and business logic - into composable elements. So those same underlying capabilities can now be built into dynamic, intelligent, open, and composable AI interfaces without rewriting the core business logic. As Salesforce Chair and CEO Marc Benioff put it in his keynote presentation, "You're going to see AI interfaces that are dynamic and intelligent, that are composable and alive. … The interface itself is alive."

Read what they're saying.

2. Purpose-trained models provide rightsized intelligence for every task.

You can certainly use a slotted screwdriver to twist in a crosshead screw, but a Phillips performs better because it's made specifically for the job. That's the thinking behind Koa, Salesforce's first purpose-built CRM reasoning model. Through a partnership with NVIDIA, Koa is built on the NVIDIA Nemotron 3 Super architecture, and it's optimized specifically for enterprise tasks.

One of the reasons Salesforce partnered with NVIDIA for this project is that Nemotron is pretrained on datasets that have clear provenance - training data that is open and available for review. "It's not a dataset that's been scraped from somewhere," said Jayesh Govindarajan, EVP of Salesforce AI and Agentforce, at a press Q&A session. "You could check what data it was pretrained on, which is really important." Then, Salesforce created synthetic simulation environments for over 14 industry personas across sales, service, marketing, and commerce to simulate real-world user interactions. This leveraged decades of Salesforce's business context on CRM workflows without using any private customer data - a level of expertise no general-purpose model can match.

As Govindarajan and Silvio Savarese, EVP and Chief Scientist of Salesforce AI Research, wrote: "We know what a service escalation looks like from the inside, what qualifying a lead actually involves, what 'resolved' means when a customer is on the other end. That … is exactly what we could give Koa."

In internal pilots, Koa proved 14% better than internal benchmarks at retaining context during complex, lengthy, multi-turn conversations. It delivered 15% more relevant answers to end users and was 11% more precise in executing the correct tool calls, meaning autonomous agents make fewer mistakes and resolve tasks faster.

Read what they're saying.

3. Slack is the front door to the enterprise.

"There's a new work operating system for the AI era, and that is Slack," said Rob Seaman, EVP and GM of Slack, at the Slack Keynote. "It's the only place where AI becomes multiplayer." In Slack, teams can work together with AI in shared channels, where agents can also connect with all the conversational context inside Slack, as well as the data, agents, and workflows inside Salesforce. Without that shared interface, AI stays stuck in "single player" mode, where one person queries a model in an isolated terminal. The individual learns; the organization doesn't. Bring agents into Slack instead, and intelligence compounds in the open.

Agents can read across channels and DMs; act on Salesforce data; and even help teams write, review, and ship code together, turning what used to be a single-player sport into a multiplayer one. Decisions become visible, reusable, and improvable by the whole organization. "Your company becomes smarter, and your company moves faster," Seaman said.

This multiplayer dynamic is already reshaping software engineering. With Slack Code, developer teams can collaborate with coding agents directly in shared channels instead of separate environments. Because the work happens in public, nontechnical staff like marketers and product managers can review the builds and guide the project alongside engineers without needing specialized developer tools. As Slack CMO Ryan Gavin explained during the session, "We are in the age of builders, and we need to turn everyone in our companies into builders."

Read what they're saying.

4. The harness and the model work together to run the Agentic Enterprise.

"We've been running our little eight-figure business with three humans and 21 agents on Salesforce for over a year," said SaaStr CEO Jason Lemkin at the keynote session, "The Enterprise AI Harness for the Agentic Enterprise." "Agents are great, right? But you can't trust them. You need a harness."

The issue isn't whether AI agents can be powerful or productive - they clearly can - but whether enterprises can safely trust them with real revenue, real customers, and real data. The AI harness is the crucial layer between raw model intelligence and the operational realities of complex organizations.

Rohan Kumar, Salesforce President and Chief Platform and Engineering Officer, laid out the six core capabilities that any enterprise harness needs to make AI trusted and governable at scale:

• Trusted models: Includes Salesforce's new CRM-tuned model, Koa, as well as frontier models like Fable so enterprises can select the right model for each use case based on price and performance.
• Trusted context: Brings together customer context with data, metadata, semantics, knowledge, real-time signals, memory, and an understanding of how work gets done across the enterprise.
• Trusted agency: Captures how agents plan and decide - the plans, steps, and runtime logic that translate a desired outcome into an executable path.
• Trusted actions: Securely connects AI to applications, APIs, workflows, tools, and business processes so agents can move from understanding what needs to happen to getting it done.
• Trusted governance: Covers data quality, classification, and policy enforcement so only appropriate information is exposed to AI.
• Trusted security: Applies identity, permissions, privacy, data protection, and runtime security to what AI can access and what agents are allowed to do.

Salesforce is pulling all of these capabilities together into one composable harness, all unified by an AI control plane to discover, manage, evaluate, and cost-govern agents across the enterprise stack.

Models alone can't run a business. They reason and can show what's probable, but they can't determine what's allowed. The Enterprise AI Harness is a way to wrap raw model intelligence in enterprise-grade context, control, and trust so organizations can safely move from experimentation with agents to running their businesses with them.

Read what they're saying.

5. ROI can happen fast.

It took one hour for Fulton Bank to go from hearing about Agentforce to setting up its first agent. Within a week, it had rolled it out to power users. Six weeks later, 3,000 employees were using it.

How? By following the three-step Agentforce ROI playbook outlined in the Agentforce Keynote, "From Fast Start to Real ROI." The playbook, said Mark Wakelin, EVP and GM for Agentforce, unfolds in three steps: Pick a use case and deploy it fast; customize and extend that use case across the enterprise; and observe and optimize agents in real time.

Fulton Bank used Hunter, a prebuilt outbound sales agent, to get going quickly. Avanthika Ramesh, Senior Director of Product Management for Agentforce, described how Hunter helped set concrete revenue goals (like reengaging $340,000 of at-risk pipeline), automatically generate and schedule tasks, and draft outreach in the seller's own voice by connecting to tools like Gmail. For Fulton Bank, this meant they could turn Hunter on and have it "literally hunting for new opportunities" almost immediately.

Southwest Airlines built a focused AI assistant to answer customer questions, freeing human reps for high-empathy interactions. Then it expanded into more complex processes like name changes, receipt requests, and travel vouchers, orchestrated via subagents within a unified "superagent" experience. Southwest achieved millions in productivity gains and seven times the ROI in a relatively short time frame.

Better still, Southwest improved its agents by observing how they were interacting with its customers. "Your customers will provide you their needs a lot faster in real time than all the planning sessions you could have had internally," said Justin Bundick, VP of AI & Intelligence Platforms at Southwest Airlines. "You do that by analyzing the conversations that are happening between the customer and that AI product and then making tweaks and changes on a very rapid basis. It's a brand-new operating model that those companies are going to have to adopt."

Got seven minutes? Watch the Dreamforce keynote highlights here.

Go deeper:

• How higher-education institutions are putting agents to work.
• Why Salesforce post-trained its own reasoning model.
• Read all the announcements from Dreamforce 2026.